Technical pieces on legal AI verification, extraction methodology, and the engineering problems that don't have easy answers. For teams building with legal data.
RAG retrieves text. It does not determine whether that text represents a binding holding, who said it, whether the case is still good law, or which courts are bound by it. These are not edge cases. They are the definition of legal verification.
Read the piece →A single accuracy number is meaningless without a methodology. Ours scores across seven independent classification categories on a human-labeled held-out set.
Coming soonA circuit split is not just a legal curiosity. It is a structured data problem: two courts, same statute, opposite holdings, both binding within their circuits. Here is how we model it.
Coming soonHohfeld's 1913 framework distinguishes duties from rights, privileges from liabilities. Modern semantic search treats all legal relationships as edges of equal weight. The consequences are concrete and computable.
Coming soonOpinion date, statute effective date, and applicability window are not the same thing. Conflating them is the most common source of temporal error in legal AI systems. We store all three, separately.
Coming soon